Detecting Online Polarization (SemEval-2026 Task 9)

Overview

Multilingual polarization detection (binary classification: Polarized vs Non-Polarized) with emphasis on fair evaluation (Macro-F1).

Problem

Online polarization is a growing societal concern, but detecting it automatically across multiple languages remains challenging due to linguistic diversity and class imbalance.

Solution

A multi-stage pipeline from TF-IDF baselines through transformer models (XLM-R, mDeBERTa-v3, RemBERT) with focal loss, weighted sampling, and soft-voting ensembles.

Highlights

• •TF-IDF + Logistic Regression baseline
• •Neural baselines: BiLSTM, BiLSTM + Attention with language identity features
• •Transformers: XLM-R (base/large), InfoXLM, mDeBERTa-v3, RemBERT
• •Imbalance strategies: focal loss, weighted sampling / inverse-frequency weighting
• •Ensembling: soft voting / weighted combination across models
• •Parameter-efficient tuning exploration (QLoRA-style) for compute constraints

Tech Stack